Most people set personal and professional goals. If you work in software, your near-term professional goals might sound like this:
- Ship the next version of our team’s product without any undiscovered crashing bugs
- Improve the performance of the caching layer by 10%
- Become a committer on some particular open-source project
Those probably sound pretty reasonable, right? They aren’t vague wishes, they’re goals. They probably aren’t totally unrealistic, and you’ll know you’ve achieved them when some objective, measurable thing happens. (They’re SMART, in other words.)
We’ll come back to these goals. First, though, I’d like digress and talk about bike racing, and specifically about the kinds of goals an amateur racer might set for a year. Maybe those would look like this:
- Stay healthy and avoid injuries all season
- Improve my one-minute power output by 0.5 W/kg
- Upgrade my racing category
These are more or less analogous to the professional goals. They’re objective, measurable, and reasonable. However, they’re also at least partially out of your control. You can’t totally prevent yourself from getting sick or crashing. Your one-minute power might be right at the limits of your genetic capabilities (or maybe your season shapes up so that it makes more sense to focus on improving other aspects of your fitness). And the category upgrade is up to an official in your local cycling association – you can get an automatic upgrade by winning races, but whether or not you win is also largely out of your control.
I first read about the idea of process and outcome goals in Joe Friel’s The Cyclist’s Training Bible, but I’ve found them extremely helpful in talking about professional goals as well. The idea is that there are two kinds of goals: outcome goals, which are results, and process goals, which relate to how you prepare and execute. Of course, you’re ultimately interested in outcomes (winning races, shipping great products, becoming recognized as an authority in your field), but outcomes are always to some extent out of your control.
Instead of focusing on outcomes for goals, it makes more sense to focus on what you can control: the processes that support those outcomes. So, in the bike case, you could recast the outcome goals as:
- Make sure I get at least 7 hours of sleep a night and avoid sketchy racers in corners or sprints
- Do plyometric exercises at least twice a week in the offseason
- Identify key races that I can do well in and make sure to show up for them well-rested and with adequate nutrition; then focus on avoiding particular tactical mistakes I’ve made in past races
The software goals I mentioned at the beginning of the article are also outcome goals. You can’t guarantee that your product won’t have any undiscovered crashing bugs, but you can sure put processes in place so that you’re more likely to discover crashing bugs before you ship. Similarly, a 10% performance improvement sounds great, but it might require unjustifiable effort or be unachievable (conversely, it might turn out that 10% improvement is insufficient to improve the behavior of the whole system)! Community leadership positions always involve political as well as meritocratic considerations; while you can make a case for yourself as a good candidate, you can’t force people to vote for you.
Recast as process goals, the software goals might look like this:
- Ensure that no new commit to our product introduces untested code, use property-based tests for all internal data structures, and incorporate fuzzing of every public interface into the QA process.
- Profile the caching layer under a range of conditions and document the hot spots, identifying algorithmic and implementation-level avenues for performance improvement. Add performance testing to CI.
- Participate in the project community, answering questions on the mailing list every week and submitting at least one nontrivial patch a week.
It’s easy to imagine the rest of your career and think about outcomes you want to see: earning a prestigious position, achieving acknowledged impact in your department or company, or writing a popular framework (or book).1 But those goals are often pretty far off from what you need to do to get there. Setting process goals supports the outcomes that you want to achieve. But the really remarkable thing is that they do so much more than that. By focusing on improving your preparation and execution, you’re making yourself a better engineer (or bike racer) even if you don’t get the outcomes you want right away – and making it more likely you’ll see great outcomes in the future.
Outcome goals are a particularly dangerous trap if you’re coming from an environment like academia or open-source development, where outcome-related metrics like tenure, citation count, and number of GitHub “stars” are never far from your evaluation of your own career. (The frictionless recognitions afforded by social media “likes” and “favorites” bring this headache to a much broader audience.) ↩
Elasticsearch has offered Hadoop
OutputFormat implementations for quite some time. These made it possible to process Elasticsearch indices with Spark just as you would any other Hadoop data source. Here’s an example of this in action, taken from Elastic’s documentation:
However, the latest versions of the Elasticsearch client offers more idiomatic Spark support to generate RDDs (and Spark data frames) directly from Elasticsearch indices without explicitly dealing with the Hadoop formatting machinery. This support is extremely cool, but the documentation is still a bit sparse and I hit a few snags using it. In this post, I’ll cover the problems I encountered and explain how to make it work.
Native Spark support is available in
elasticsearch-hadoop version 2.1.0. If you’d tried to set this up last week, you’d have needed to run against a snapshot build to get support for Spark data frames. Fortunately, the most recent beta release of
elasticsearch-hadoop includes this support. Add the following library dependency to your project:
Since we won’t be using the lower-level Hadoop support, that’s the only library we’ll need.
We need to supply some configuration in order to create RDDs directly from indices.
elasticsearch-spark expects this configuration to be available as parameters on our Spark context object. Here’s an example of how you’d do that in a vanilla Spark application:
Note that since the default value for
localhost, you don’t need to set it at all if you’re running against a local Elasticsearch instance. If you were running Elasticsearch behind a reverse proxy running on
proxyhost on port 80, you might specify
proxyhost:80 instead (and you would probably also want to set
false so your client wouldn’t discover nodes that weren’t reachable outside of the private network). There are other options you can set, but this is the most critical.
elasticsearch-spark offers implicit conversions to make RDDs directly from Elasticsearch resources. (Careful readers may have already seen one in action!) The following code example assumes that you’ve initialized
spark as a
SparkContext with whatever Elasticsearch configuration data you need to set:
This is far more convenient than using the Hadoop
InputFormat, but the interface still leaves a little to be desired. Our data may have a sensible schema, but the result we’re getting back is as general as possible: just a collection of key-value pairs, where the keys are strings and the values are maps from strings to arbitrary (and arbitrarily-typed) values. If we want more explicit structure in our data, we’ll need to bring it in to Spark in a different way.
Generating data frames
In order to benefit from the structure in our data, we can generate data frames that store schema information. The most recent release of
elasticsearch-spark supports doing this directly. Here’s what that looks like, again assuming that
spark is an appropriately-configured
Once we have a data frame for our ES resource, we can run queries against it in Spark:
Elasticsearch’s Spark support offers many additional capabilities, including storing Spark RDDs to Elasticsearch and generating RDDs from queries. Check out the upstream documentation for more!
Silex is a small library of helper code intended to make it easier to build real-world Spark applications;1 most of it is factored out from applications we’ve developed internally at Red Hat. We have a couple of long-term goals for the Silex project:
- We want to make it very easy for us to spin up on new data-processing problems without spending a lot of time dealing with accidental complexity, and
- we want to have a generally-useful standard library atop Spark that provides primitives and solutions to simplify common tasks and reflects best practices as we discover them.
Part of having a library that people will want to use is having a quality test suite and continuous integration. There are some tricks to developing automated tests for Spark applications, but even keeping those in mind may not be sufficient to let you test your Spark apps in a hosted CI environment. This post will show you how to bridge the gap.
Continuous integration challenges
Since we want to make it easy to review contributions to Silex, I set up Travis CI to watch branches and pull requests. Travis CI is a great service, but if you’re used to running tests locally, you might have some problems running Spark-based tests under hosted CI environments. Here’s what we learned:
- When running tests locally, creating a Spark context with
local[*](in order to use all of the available cores on your machine) might be a good idea. However, a hosted CI environment may offer you a lot of cores but relatively little memory, so your tests might be killed because each Spark executor has nontrivial memory overhead. Instead, consider limiting your code to use fewer cores.
- If you’re using Kryo serialization and have set the
spark.kryoserializer.buffer.mbproperty to something large (perhaps because you often have serialization buffer crashes in production), don’t be surprised if you run out of memory while running in CI. Spark doesn’t share serializers between threads, so you could be allocating a huge buffer for each thread even if your test code doesn’t need to serialize anything all that large.2
- Spark SQL (at least as of version 1.3.1) defaults to creating 200 partitions for shuffles. This is probably a good starting place for real-world data, but it’s overkill for test cases running on a single machine. Furthermore, since each partition has some extra memory overhead, it’s another possible cause of OOMEs.
Fixing the problems
Silex provides an application skeleton trait, and we use a class extending this trait to package up Spark and SQL contexts for test cases. That class isn’t currently part of the public Silex API, but you can see what it looks like here:
As you can see, we allow users to specify a Spark master URL but default to using two cores locally. Furthermore, we use the
addConfig function from
AppCommon – which takes a
SparkConf and returns a modified
SparkConf – to ensure that our Kryo buffer size is the Spark default of 2 mb, rather than the larger Silex default.
If you’re used to writing test code that exercises Spark, you probably already have boilerplate (using something like ScalaTest’s
BeforeAndAfterEach) to set up and tear down a Spark context for each test case. We set the Spark SQL property to control parallelism in data frame and SQL shuffles in the test setup code itself:
Good CI is one of those things that makes it fun to develop software, but the accidental differences between a hosted environment and your local environment can be totally frustrating. By taking into account the constraints you’re likely to see when running in a resource-limited container on a CI server, you can help to ensure that your tests only fail when your code is actually broken.
Natural join is a useful special case of the relational join operation (and is extremely common when denormalizing data pulled in from a relational database). Spark’s DataFrame API provides an expressive way to specify arbitrary joins, but it would be nice to have some machinery to make the simple case of natural join as easy as possible. Here’s what a natural join needs to do:
- For relations R and S, identify the columns they have in common, say c1 and c2;
- join R and S on the condition that R.c1 == S.c1 and R.c2 == S.c2; and
- project away the duplicated columns.
so, in Spark, a natural join would look like this:
We can generalize this as follows (note that joining two frames with no columns in common will produce an empty frame):
Furthermore, we can make this operation available to any
DataFrame via implicit conversions:
One of the great things about Apache Spark is that you can experiment with new analyses interactively. In the past, I’ve used the
sbt console to try out new data transformations and models; the console is especially convenient since you can set it up as a custom Scala REPL with your libraries loaded and some test fixtures already created.
However, some of Spark’s coolest new functionality depends on some aspects of Scala reflection that aren’t compatible with how
sbt uses classloaders for tasks, so you’re liable to see
MissingRequirementError exceptions when you try and run code that exercises parts of Spark SQL or the DataFrame API from the
You can certainly run a regular Scala REPL or the
spark-shell, but doing so sacrifices a lot of the flexibility of running from
sbt: every time your code or dependencies change, you’ll need to package your application and set your classpath, ensuring that all of your application classes and dependencies are available to the REPL application.
Fortunately, there’s an easier way: you can make a small application that runs a Scala REPL set up the way you like and ask
sbt how to set its classpath. First, write up a simple custom Scala REPL, like this one:
ReplApp application sets up a Scala REPL with imports for some common Spark classes and bindings to a
ConsoleApp object is just a simple wrapper for Spark context and configuration; see the Silex project, where my team is collecting and generalizing infrastructure code from Spark applications, for more details – or just change this code to set up a
SparkContext as you see fit.)
In order to run this application, you’ll need to set its classpath, and
sbt gives you a way to do find out exactly what environment it would be using so you can run the application manually.2 First, make sure you have a copy of
sbt-extras either in your repository or somewhere pointed to by
SBT in your environment. Then, create a shell script that looks like this:
You can then run
repl.sh and get a Scala REPL that has all of your app’s dependencies and classes loaded and will let you experiment with structured data manipulation in Spark.
Frustratingly, apps that use these features will work, since the classes Scala reflection depends on will be loaded by the bootstrap classloader, and test cases will work as long as you have
sbtfork a new JVM to execute them! Unfortunately,
sbtcannot currently fork a new JVM to run a console. ↩
run-maintask is the right way to run most applications from
sbt, but it seems to be somewhat flaky when launching interactive console applications. ↩
Last night I had a crazy realization: I could probably replace the majority of what my team hopes to accomplish with standup meetings, design documents, project management apps, and social code sharing features with a single tool.
Before I reveal what this tool is, let’s consider why most team collaboration is suboptimal. Everyone who has been in standup meetings or read groupwide status reports has probably seen the following status-update antipatterns; in fact, unless you’re a better teammate than I, you’ve probably been guilty of each of these yourself:
- “This week I read 253 emails in my inbox, 1478 emails from mailing list folders, and sampled 75 messages from my spam folder to make sure none of them were spuriously filed. I also wrote a net 498 lines of code, brewed eight pots of coffee and five shots of espresso, spent 246 minutes in meetings, and urinated twenty-one times.”
- “I’m still working on the same thing I was working on last week. None of you are close enough to this problem to know why it is so tough and I’m too defeated to explain it to you, so feel free to assume that I am either heroically deep in the weeds or have gotten totally ratholed on something irrelevant.”
- “I discovered that this graph search problem could be better modeled as a language-recognition query, and spent some time investigating efficient representations before coming up with a really clever approach involving boolean satisfiability – there was a paper in the SIGYAWN ‘75 proceedings that used a similar technique; maybe you read it too – … [fifteen minutes pass] … and improved performance by 0.57%.”
It doesn’t seem like it should be, but getting the level of detail right in status updates is hard, and it’s hard for stupid reasons. Your big accomplishments for the week are easy to explain and you don’t want to make it seem like you did nothing, so you start accounting for every second since last Monday. It’s too depressing to consider that the bug that’s been keeping you awake for the last few weeks still isn’t solved and you don’t want to spend time talking about all of the dim trails you investigated, so you’ll Eeyore it up to get out of the meeting and back to work. You’re so deep in a problem that you think everyone else needs to know about a bunch of tiny details and can’t tell that they’ve all fallen asleep.
To some extent, all of these seemingly-different problems stem from too much status and not enough context. But even if we recast status reports as context updates, we’ve traded the problem of finding an appropriate level of status detail for finding an appropriate amount of context. How do we decide what an appropriate amount of context is?
Think of the best and worst talks you’ve ever seen. A good talk doesn’t exhaustively describe the work the author did, but it places it in enough context so that the audience has some way to evaluate the claims, get interested, and decide if the work is worth investigating further. A bad talk is vague, disorganized, gets bogged down in irrelevant details, or only makes sense to people who already know everything about its subject.
Now think about the best talks you’ve ever given. Giving a good talk – even more so than writing a good article or essay – hones and clarifies your argument. You have to step out of your bubble and see your work from the perspective of your audience, because you have just a few minutes to catch their attention before they tune out and start checking their email while waiting for the next speaker. Now, I’ve given a lot of lectures, academic talks, and conference presentations, but I’ve found that giving informal, internal tech talks on some interesting aspect of something I’ve been working on has paid off disproportionately.
So I don’t want my team to give each other status updates. Instead, I want us to give regular short talks. And the medium that I want us to use as we develop these talks is a shared slide deck, updated throughout the week.1 In spite of the amazingly persistent ubiquity of speakers reading through their wall-of-text decks like apparatchiks in dystopian fiction, slides aren’t a substitute for talks. However, a good slide deck supports a good talk, and designing a slide deck helps to make sure that you have the right amount of context in your talk, because the wrong amount of context in a talk leaves some obvious odors in the deck:
- Oh, you need to drop down to 18 point type to cover everything you wanted here? That’s probably a clue that you’re covering something that would be better addressed in further discussion or a subsequent deep dive.
- Does this slide feel totally bare? Maybe you need more background here.
- Could you draw a picture and replace this paragraph? Could you write a sentence and replace this code excerpt?
- If you had to explain this to a customer or an executive – really, to anyone who’s sharp and engaged but not a domain expert – right now, where would you start? How would you keep them interested?
My team is doing data science, so a lot of what we wind up delivering turns out to be presentations and visual explanations. As a result, this practice could be even more beneficial for us than it might be for other groups. Rigorously working out explanations among ourselves is certain to pay off in the future, when we have to present results internally or externally. By using a shared deck, we can annotate with questions, link to more detail elsewhere, and rapidly iterate on our presentations as we have improved results or come up with clearer explanations.
(Thanks to RJ Nowling, who suggested the idea of a regular “analytics review,” like a code review, and got me thinking along these lines.)
Over eight years ago, Richard WM Jones wrote a great but disheartening article about his experience serving as a technical reviewer for an infamous book about OCaml. The post made quite an impression on me at the time and I’ve often recalled it over the years whenever opportunities to do prepress reviews have landed in my inbox. Briefly, Jones was asked to use terrible tools (Microsoft Word) to deal with stilted, error-ridden prose surrounding unidiomatic and broken code. For his trouble, he got an hourly wage that would have represented a slight raise over what I made mowing my neighbors’ lawns in high school.
Recently, I was invited to review a book under even more outrageous terms. The project in question was the second edition of a book on a subject I understand pretty well.1 I didn’t respond to their first invitation to review because:
- I assumed they’d cast a fairly wide net;
- my spare time is pretty well saturated these days and reviewing technical prose is a lot of slow, precise work;
- in my opinion, this publisher releases a lot of subpar products and I don’t want my name and reputation associated with a mediocre book;
- the earlier edition of this book had received extremely poor reviews; and
- the compensation offered was simply risible: credit in the frontmatter of the finished book, a paper or ebook copy of the finished book, and an additional ebook of my choice.
Now, there are always tradeoffs involved in choosing contract work: one can weigh compensation against loving one’s work, having pride in products, having a prestigious position, or enjoying a low-stress work environment. But I’m pretty sure I know what the skill set that would enable someone to be a competent reviewer of a book like this one is worth, and the people who have it aren’t likely to be motivated to volunteer for a for-profit corporation in exchange for credit and free ebooks. Personally, a demand on my spare time that depends on my professional expertise (and takes away from time I might spend with my family, bicycling, catching up on sleep, or learning something new) but offers no compensation is not particularly compelling, and it becomes far less compelling when it seems like an opportunity to be involved with a product that I’m not likely to believe in and that I probably wouldn’t be proud to put my name on. So when the publisher sent me a rather snippy message following up, I wrote back, gently indicating that I would not be able to do the review and that they’d likely have trouble finding a qualified reviewer who was willing to work for free.
One wonders what publishers expect to accomplish with prepress technical reviews. In Jones’s case, he provided a lot of feedback and suggested that Apress not proceed with publishing Practical OCaml without major revisions; they ignored his objections but used his name in the final text. (At least he was paid something, I guess.) Not every publisher works this way, but the ones that do seem more interested in shifting the blame for bogus content to an external contractor than in releasing a thoroughly vetted product.
Requests to do specialized work for free are almost always insulting. However, the apparent value that some publishers ascribe to prepress review is even more insulting if you consider it from the perspective of the technical-book consumer!
In fact, this publisher had invited me to write a book for them on this topic earlier this year, apparently before deciding to revise their existing offering instead. ↩
I’ll be speaking later this afternoon at ApacheCon EU. The title of my talk is “Iteratively Improving Spark Application Performance.” The great thing about Apache Spark is that simple prototype applications are very easy to develop, and even a first attempt at realizing a new analysis will usually work well enough so that it’s not frustrating to evaluate it on real data. However, simple prototypes can often exhibit performance problems that aren’t obvious until you know where to look.
In this talk, we’ll introduce Spark’s execution model and discuss how to use Spark effectively. I’ll use a prototype implementation of my bike data analytics application as a running example and will present four general principles to keep in mind when writing efficient Spark applications, complete with detailed code explanations. My slides are available here as a PDF.
If you’re interested in more detail about the bike data analytics application, you can watch a brief video demo or watch a video of my talk about this application from Spark Summit earlier this year. Finally, I’ve published a blog post covering similar principles for improving Spark applications, which will be a useful reference whether or not you’re able to attend the talk.1
If you’re also at ApacheCon today, I hope to see you at 15:50 CET!
The talk will feature visual explanations and other material that are not in that post, but the post has links to full code examples suitable for off-line experimentation. ↩
My last post covered some considerations for using Spark SQL on a real-world JSON dataset. In particular, schema inference can suffer when you’re ingesting a dataset of heterogeneous objects. In this post, I’d like to sketch out some ways to connect schema inference to type inference, in order to point to automated solutions to some of the problems we’ve seen.
We’ll start out by reviewing some definitions from the world of type systems:
- a type is a range of values.
- a product type is a compound type that describes structures with multiple values of (possibly) different types (e.g., tuples or records). You can think of a product type as being like the Cartesian product of several other types: if we have types A and B, the type A × B (i.e., the type of pairs of one value in A and one value in B) can represent the Cartesian product of A and B.
- a sum type is a compound type that represents the disjoint union of several other types. Given types A and B, the type A + B represents the sum of A and B (i.e., any value in A or any value in B).
In Scala terms, a simple example of a product type is a tuple:
and a simple example of a sum type is a basic linked list:
Note that, in the case of the linked list (as with other standard recursive data structures), we’re actually looking at a sum type in which one of the summed types is a product type: an
IntList can either be an
EmptyList or a cons cell, consisting of a pair of an
Int value (the head of the list) and an
IntList (the tail of the list).
Also note that, in Scala, sum types are tagged. You can statically distinguish between cons cells and the empty list in a case match.1 C programmers are familiar with a related kind of sum type: the untagged union, which describes a structure that can hold one value that can be of several different types. If you have a C
union, you’ll need to explicitly track what kind of value it’s holding since you won’t be able to tell what kind of value it contains statically.
Given a collection of JSON objects,2 a schema inference algorithm produces a type that is the supertype of every object in the collection. (This is an interesting problem only if the collection contains objects of different types.) As an example, consider the following two objects:
The first models some characteristics of a bicycle and the second models some characteristics of a car. Both have
gears, but bicycles have data about what kind of
handlebars they include; cars have data about whether or not they include anti-lock brakes (
abs) and specify what kind of
fuel they expect. If we ask Spark SQL to infer a schema for a collection including these two objects, we’ll get something like this:
root |-- abs: boolean (nullable = true) |-- fuel: string (nullable = true) |-- gears: integer (nullable = true) |-- handlebars: string (nullable = true) |-- pedals: string (nullable = true)
Given a collection of objects with types T₁, T₂, ⋯, Tₓ, what we’re trying to get is the sum type T₁ + T₂ + ⋯ + Tₓ . The way we get this type is as an untagged union: we get an object that has the union of all of the fields of T₁, all of the fields of T₂, and so on. (Put another way, we’re really treating a heterogeneous collection as a union of homogeneous collections and identifying its schema as the union of the schemas for each included homogeneous collection.) Just as if we were using a C
union, we’d have to figure out whether or not each object was a bicycle or a car. Fortunately, since we’re be in a safe-language context, we’d know that an object with a value in the
pedals field wouldn’t have a value in the
It should be clear that this is a totally reasonable approach, since it combines a straightforward implementation with a pleasant and unsurprising user experience: by adding a simple filter to my queries, I can select only objects of a given type from a heterogeneous collection. For example, if I want to find the mean number of gears for every bicycle with drop handlebars, I could write the following query (assuming I’d registered my
SchemaRDD as a table called
handlebars will be
NULL for every car, this query will only consider bicycles. Other collections may not admit such straightforward differentiation between object kinds, of course.
Problems and possible solutions
Tagging diverging types
As I mentioned last week, things begin to break down when we attempt to construct the sum of types that have the same field names but diverging types for those fields, as in the following example:
A few days ago, Spark would have failed to infer a schema for this collection of objects, but open-source development moves quickly and now it will handle the conflicting types by giving users a schema in which the
branches field contains an array of strings (and in which every object in the collection has its
branches field cast to an array of strings).
In general, it seems like whatever technique we might adopt schema inference should prefer imprecise results to a crash. But are there ways Spark could do a better job of inferring schemas for heterogeneous collections with diverging field types? I believe there might be, and part of the answer is going from untagged unions to tagged unions.
The fedmsg data set is a heterogeneous collection of objects; in some of these objects, the
branches field models one concept as an array of strings, and in others, it models a different concept as an array of objects. I solved this problem by preprocessing the JSON to rename the different
branches fields – in effect, dividing the
branches fields into two tagged possibilities, like this. We could imagine that Spark could handle schema divergence by automatically identifying equivalence classes of field types, perhaps naming them with an integer subscript (e.g.,
branches_2) or with some abbreviated identifier for the value type, as follows:
Here’s an alternative way we could make the tagging explicit, by wrapping the diverging field’s values in an object:
The advantage to this tagged-union approach is that it makes it possible to write queries that handle heterogeneous collections. The disadvantage is that the path from the schema root to the tagged value will be different from the path to that value in the schemas of any of the objects in the original heterogeneous collection. The schema change may be surprising to the user, but – by definition – any approach to derive a fixed schema for a collection of objects with diverging schemas will not result in a schema that is totally compatible with the objects in the original collection.
The other problem I observed related to encoding maps as JSON objects, as in the following example where we have a map from bicycle names to vehicle objects:
The inferred schema for a collection including an object like this will have a distinct field for every key in every map. In this case, we don’t have an immediately great solution from the broader world of type systems. However, it seems that some heuristic – for example, “if an object has more than n fields and each has a value of the same type, it is a map” – might be able to do a good job of identifying objects that model maps.
In any case, it is possible for the user to preprocess their objects in order to encode maps in a different way if that will make it easier to query the data after ingest.
Spark SQL’s support for schema inference is an extremely cool feature, although it obviously isn’t able to work magic by ascribing rigid structure to heterogeneous collections of ad hoc objects. Some minor improvements could make inferred schemas more useful for these kinds of collections, though.
Scala doesn’t support untagged union types directly – although there are clever ways to encode them in the type system – and Scala won’t actually be able to infer sum types from collections of untagged values: the inferred type of
Array[String + Int]. ↩
(Slightly) more precisely: objects are defined recursively as unordered collections of key-value pairs, where keys are strings and values can be numbers, booleans, strings, objects, or arrays of values. ↩
In this post, I’ll briefly introduce fedmsg, the federated message bus developed as part of the Fedora project’s infrastructure, and discuss how to ingest fedmsg data for processing with Spark SQL. While I hope you’ll find the analytic possibilities of fedmsg data as interesting as I do, this post will also cover some possible pitfalls you might run into while ingesting complex JSON data or the contents of large SQL databases into Spark SQL more generally.
fedmsg is one of the coolest things about Fedora’s infrastructure.1 Because nearly everything that happens in Fedora is logged to a messaging bus, there is a vast amount of public data about the daily operation of a large and thriving open-source community, all in one place.2 The whole list of message topics published on the Fedora project’s fedmsg bus are documented here, but these examples should give you a taste of the kinds of messages available for inspection:
- when a Fedora contributor updates the package build metadata (including RPM specfile, patches, source tarball list. etc.) in
- when a package build is started (and then again when the package build succeeds or fails, and when the package build is pushed out to the distribution),
- when a Fedora community member tags a package with keywords,
- when a blog post is syndicated to Fedora Planet,
- when a wiki page is created or updated,
- when an IRC meeting starts or ends,
and so on. Furthermore, these data are accessible either as a stream of messages (as they are generated), as JSON-encoded historical data via a REST interface, or as a PostgreSQL dump of all historical fedmsg activity.
An exhaustive treatment of fedmsg is outside the scope of this post,4 but I wanted to make it clear that
- open infrastructure for a very large open-source project is an extremely cool thing,
- there is a lot of interesting data available in fedmsg,
- there have been some cool applications built atop fedmsg already (see Fedora Badges for an Xbox-like “achievements” system or Ralph Bean’s animated visualization for a couple of fun examples), and
- there are a lot of possibilities for making sense of this data using Spark, Spark SQL, and Spark Streaming.
Getting the historical data
I wanted to look at a large amount of data to start with, so I began by grabbing a dump of the datanommer database. (Datanommer is a service that records each message on the fedmsg bus in a SQL database.) The nightly dump I got was about 3 gb compressed and expanded to about 40 gb – certainly not “big data,” but too big to fit in main memory on my workstation. I loaded the dump into a new Postgres database and started inspecting it.
The largest table by far is
messages,5 which has the following schema:
There are a couple of things to note here:
signaturefields account for the bulk of the raw data, and
_msgfield actually contains the string representation of a topic-specific JSON object.
As we’ll see, each of these can lead to some pitfalls when attempting to ingest the data into Spark.
Pitfall 1: Single-node memory limits and Slick projections
Here’s the core of a very simple program that ingests fedmsg records from a local PostgreSQL database, using Slick, and outputs these to JSON files,6 one for each record,7 in a local directory (I’m not reproducing the generated code that models datanommer tables and rows here):
Note that the
message2json implicit conversion doesn’t do anything with the
These fields are necessary for critical production applications because anyone can write to the fedmsg bus. However, I’m not planning to make any irreversible decisions based on my analyses of fedmsg data and am thus not currently interested in actually verifying message contents or provenance. Since I’m also not currently interested in wasting memory or performance, I’m going to strip these out of the generated JSON.
Unfortunately, even shipping these fields from the database to the JVM only to ignore them dramatically impacts the performance and memory footprint of our importer, so we’ll want to project these away sooner. The best solution I found for efficiently projecting away nullable fields from a SQL database accessed via Slick involved modifying the generated code to control how case classes were instantiated from the database. The generated code includes definitions of case classes for rows in each table, like this:
The generated code also contains classes describing each table; these include a
* method defining the default projection from a tuple to a row-specific case class. In our case, the default projection for the
messages table looks like this:
If we modify it, we can ignore
signature (essentially returning SQL
NULL values instead of enormous encoded objects for these fields):
Now we can fire up a REPL and produce JSON documents from the datanommer data; the following example will take the first 100 rows PostgreSQL gives us and generate documents for them in
Pitfall 2: The Lost Caverns (of type inference)
Spark SQL supports Hive-style records and arrays within relational data. An especially cool feature of Spark SQL, though, is that it can infer a HiveQL schema from a collection of JSON documents with complex nested structure (see the
jsonFile method of
SQLContext, which will operate on a single JSON document or a directory of JSON documents).
If you have a clone of the Spark repository on your machine, the easiest way to try this out is to launch
sbt hive/console from within your local clone. You’ll then be able to build a schema from some JSON documents:
JSON is convenient for users because it lets us specify untyped objects with no fixed schema. Unfortunately, JSON lets us specify untyped objects with no fixed schema. This means that Spark might have trouble inferring a fixed schema for our data. Consider how type inference works in real programming languages: if a variable can take one of two possible values, its type must be the unification on their types. Sometimes, this means we get a trivial typing (in Scala terms,
Any), but in other cases, it means we get a contradiction (in Scala terms,
Nothing). When Spark is inferring types for our documents, it will simply raise an exception if it tries to unify two objects with diverging types, as you can see in the following example:
The two JSON objects each contain a field called
branches, but in the former it is an array of strings, while in the latter it is an object. In the case of fedmsg data, this problem came up because different message topics used
branches to correspond to different (but related) concepts. If we could rename these different cases, we could avoid the diverging types. Since the JSON library we’re using has a facility to transform JSON values based on partial functions, we can do this easily:
renameBranches method above will simply translate fields named
branches with array-of-strings values to fields named
pkg_branches with the same values and translate fields named
branches with array-of-objects values to fields named
git_branches with the same values. Other fields are ignored by this code.
Pitfall 3: Too many columns
A related problem comes up when we have JSON objects who use field names as entity names, like in the following example (taken from a message describing a
files is essentially a collection of objects describing different changed files in the commit (
sources), but it is represented as an object with a field for each changed file. Because of this, the inferred schema will have a field in the
stats.files object for every single filename that has ever been committed to Fedora’s
git repository! (That’s a lot of
NULL values in a typical row.)
There are other similar cases in the fedmsg data as well where entity names become field names; these can all be handled by transformation rules in fairly straightforward fashion.
This post should have enough background to help you get started looking at fedmsg data with Spark SQL. In future posts, I’ll discuss some additional correctness and efficiency considerations; demonstrate how to unify warehoused, batched, and streaming data; and introduce some analytic applications involving fedmsg data.
I’ve recently been reading Jay Kreps’ I ♥ Logs, in which Kreps argues (from principle and from experience) for the centrality of the log abstraction in distributed systems. I may write a longer review later, but the argument is sound (and consonant with some of the arguments initially made for basing Fedora’s infrastructure around a messaging bus) and the book is definitely worth checking out! ↩
There are also tables to store the names of users, the names of packages, and to model the relationships between users or packages and fedmsg messages involving those users or packages. ↩
I’m using JSON here because of the rich structure of the topic-specific messages. If you were ingesting flat relational data, it would almost certainly make more sense to use JdbcRDD and do the ingest within Spark. ↩
Please don’t use the above code in production; it’s intended to be simple, not efficient. Producing one file per record makes it easy to see where things are going wrong, but it’s very slow both for this program and for the Spark jobs that you’ll want to write to read these data. Your local filesystem probably won’t be too happy with you, either. ↩